xref: /illumos-gate/usr/src/lib/libm/common/m9x/fmaxf.c (revision b515258426fed6c7311fd3f1dea697cfbd4085c6)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  */
25 /*
26  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  */
29 
30 #if defined(ELFOBJ)
31 #pragma weak fmaxf = __fmaxf
32 #endif
33 
34 /*
35  * fmax(x,y) returns the larger of x and y.  If just one of the
36  * arguments is NaN, fmax returns the other argument.  If both
37  * arguments are NaN, fmax returns NaN (ideally, one of the
38  * argument NaNs).
39  *
40  * C99 does not require that fmax(-0,+0) = fmax(+0,-0) = +0, but
41  * ideally fmax should satisfy this.
42  *
43  * C99 makes no mention of exceptions for fmax.  I suppose ideally
44  * either fmax never raises any exceptions or else it raises the
45  * invalid operation exception if and only if some argument is a
46  * signaling NaN.  In the former case, fmax should always return
47  * one of its arguments.  In the latter, fmax shouldn't return a
48  * signaling NaN, although when both arguments are signaling NaNs,
49  * this ideal is at odds with the stipulation that fmax should
50  * always return one of its arguments.
51  *
52  * Commutativity of fmax follows from the properties listed above
53  * except when both arguments are NaN.  In that case, fmax may be
54  * declared commutative by fiat because there is no portable way
55  * to tell different NaNs apart.  Ideally fmax would be truly com-
56  * mutative for all arguments.
57  *
58  * On SPARC V8, fmax must involve tests and branches.  Ideally,
59  * an implementation on SPARC V9 should avoid branching, using
60  * conditional moves instead where necessary, and be as efficient
61  * as possible in its use of other resources.
62  *
63  * It appears to be impossible to attain all of the aforementioned
64  * ideals simultaneously.  The implementation below satisfies the
65  * following (on SPARC):
66  *
67  * 1. fmax(x,y) returns the larger of x and y if neither x nor y
68  *    is NaN and the non-NaN argument if just one of x or y is NaN.
69  *    If both x and y are NaN, fmax(x,y) returns x unchanged.
70  * 2. fmax(-0,+0) = fmax(+0,-0) = +0.
71  * 3. If either argument is a signaling NaN, fmax raises the invalid
72  *    operation exception.  Otherwise, it raises no exceptions.
73  */
74 
75 #include "libm.h"	/* for isgreaterequal macro */
76 
77 float
78 __fmaxf(float x, float y) {
79 	/*
80 	 * On SPARC v8plus/v9, this could be implemented as follows
81 	 * (assuming %f0 = x, %f1 = y, return value left in %f0):
82 	 *
83 	 * fcmps	%fcc0,%f1,%f1
84 	 * fmovsu	%fcc0,%f0,%f1
85 	 * fcmps	%fcc0,%f0,%f1
86 	 * fmovsul	%fcc0,%f1,%f0
87 	 * st		%f0,[x]
88 	 * st		%f1,[y]
89 	 * ld		[x],%l0
90 	 * ld		[y],%l1
91 	 * and		%l0,%l1,%l2
92 	 * sethi	%hi(0x80000000),%l3
93 	 * andn		%l3,%l2,%l2
94 	 * andn		%l0,%l2,%l0
95 	 * st		%l0,[x]
96 	 * ld		[x],%f0
97 	 *
98 	 * If VIS instructions are available, use this code instead:
99 	 *
100 	 * fcmps	%fcc0,%f1,%f1
101 	 * fmovsu	%fcc0,%f0,%f1
102 	 * fcmps	%fcc0,%f0,%f1
103 	 * fmovsul	%fcc0,%f1,%f0
104 	 * fands	%f0,%f1,%f2
105 	 * fzeros	%f3
106 	 * fnegs	%f3,%f3
107 	 * fandnot2s %f3,%f2,%f2
108 	 * fandnot2s %f0,%f2,%f0
109 	 *
110 	 * If VIS 3.0 instructions are available, use this:
111 	 *
112 	 * flcmps	%fcc0,%f0,%f1
113 	 * fmovslg	%fcc0,%f1,%f0	! move if %fcc0 is 1 or 2
114 	 */
115 
116 	union {
117 		unsigned i;
118 		float f;
119 	} xx, yy;
120 	unsigned s;
121 
122 	/* if y is nan, replace it by x */
123 	if (y != y)
124 		y = x;
125 
126 	/* if x is nan, replace it by y */
127 	if (x != x)
128 		x = y;
129 
130 	/* At this point, x and y are either both numeric, or both NaN */
131 	if (!isnan(x) && !isgreaterequal(x, y))
132 		x = y;
133 
134 	/*
135 	 * clear the sign of the result if either x or y has its sign clear
136 	 */
137 	xx.f = x;
138 	yy.f = y;
139 	s = ~(xx.i & yy.i) & 0x80000000;
140 	xx.i &= ~s;
141 
142 	return (xx.f);
143 }
144